| Wheat stripe rust, caused by Puccinia striiformis f. sp. tritici (Pst), is one of the important and destructive crop diseases worldwide. The early monitoring and quantitative detection of wheat stripe rust has great significance to the disease prevalence, the rational uses of pesticides, and the quality safety and the yield safety of wheat. In this study, the hyperspectral data were collected from wheat canopy during the latent period inoculated with three different concentrations of urediniospores and classification models based on three different methods were built to differentiate leaves with and without infection of the stripe rust pathogen. With the different Pst races, the hyperspectral data were collected from wheat leaves canopy during the latent period to assess the stripe rust pathogen exist or not. A new method to obtain spectral refletance, the blackboard method was used to test the models accuracy and the black background influence to the modeling. The macro and micro system was constructed to assess the wheat stripe rust in latent period in two years field experiments. The results were as follows:1. It is practicable to use the hyperspectral reflectance combined with the Pst concentration to detect the latent period wheat stripe rust in field based on DPLS and SVM methods.In the 325-1075 nm waveband:The average accuracy rate of the models based on DPLS or SVM was between 75-80%, the optimal models average accuracy rate was between 80-85%. In the subband: the models based on DPLS, the better performance models was focused on the waveband of 325-474 nm and the spectral feature of Reflectance in the two years field experiments; The optimal models in the two years were both focused on the waveband of 325-474 nm, the spectral feature of Reflectance and the samples modeling proportion of 4:1. In the models based on SVM, the better performance models was focused on the waveband of 475-624 nm and the spectral feature of the 1st derivative of Pseudo-absorption index in the two years field experiments; The optimal models in the two years were both focused on the waveband of 325-474 nm, the samples modeling proportion of 4:1 and the spectral feature of Reflectance or the derivative transmation of Reflectance.2. It is available to use the canopy hyperspectral reflectance to differentiate the different Pst concentrations in lantent period. And the effects of different inoculation days, different spectra features, wavebands and the number of the samples used in modeling on the performances of the models were assessed. The results showed that in the spectral region of 325-1075 nm, the models accuracy based on ANN method was better than the DPLS, and the SVM was the best. The model with the spectral feature of 1st derivative of Reflectance had better accuracy than others. In the subband, the models accuracy based on DPLS method was better than the ANN, and the SVM was the best. The model with the spectral feature of derivative transformation had better accuracy than others, the models with the waveband of 325-474 nm and 925-1075 nm had better accuracy than other models. After the comprehensive comparison of the models accuracy based on the three methods:the models based on SVM performed better accuracy, robustness and generalization ability, and the accuracy rate could be up to 100%.3. It is feasible to use the canopy hyperspectral reflectance of wheat stripe rust in latent period to differentiate the different Pst races. And the effects of different inoculation days, different spectra features, wavebands and the number of the samples used in modeling on the performances of the models were assessed. In the 325-1075 nm waveband, the model with the spectral feature of Reflectance had better accuracy than others. The average accuracy rate was 95.67% for the training set and 94.37% for the testing set. In the subband, the models with the waveband of 775-924 nm and the spectral feature of Reflectance or the Pseudo-absorption index had better accuracy than others, and the test set average accuracy rate could up to 96.59%.4. It is achievable to use the new method-blackboard method to assess the wheat stripe rust in latent period. In the experiment of the different Pst concentrations:In the 325-1075 nm waveband, the model with the spectral feature of the Pseudo-absorption index had better accuracy than others. The average accuracy rate was 96.13% for the training set and 93.31% for the testing set. The optimal model was the model with the spectral feature of the 2nd derivative of Pseudo-absorption index and the samples modeling proportion of 4:1. The number of principal components was 8, the training set accuracy rate was 99.17%, the testing set accuracy rate was 96.67%. In the subband, the models with the waveband of 475-624 nm and the spectral feature of the Pseudo-absorption index had better accuracy than others, the test set average accuracy rate could up to 91.48%. There were three optimal models. The First one was the model with the waveband of 775-924 nm, the spectral feature of Reflectance and the samples modeling proportion of 4:1. The number of principal components was 4, the training set accuracy rate was 100%, and the testing set accuracy rate was 100%. The Second one was the model with the waveband of 325-474 nm, the spectral feature of the 1st derivative Pseudo-absorption index and the samples modeling proportion of 4:1. The number of principal components was 8, the training set accuracy rate was 96.67%, the testing set accuracy rate was 100%. The Third one was the model with the waveband of 325-474 nm, the spectral feature of the 2nd derivative Pseudo-absorption index and the samples modeling proportion of 4:1. The number of principal components was 11, and the training set accuracy rate was 97.50%, and the testing set accuracy rate was 100%.In the experiment of the different Pst races:In the 325-1075 nm waveband, the model with the spectral feature of the 2nd derivative of Reflectance had better accuracy than others. The average accuracy rate was 96.13% for the training set and 94.71% for the testing set. The optimal model was the model with the spectral feature of the 2nd derivative of Reflectance and the samples modeling proportion of 3:1. The number of principal components was 10, the training set accuracy rate was 97.58%, the testing set accuracy rate was 98.18%. In the subband, the models with the waveband of 925-1075 nm and the spectral feature of the 2nd derivative of Reflectance had better accuracy than others, the test set average accuracy rate could up to 96.30%. There had two optimal models. The First one was the model with the waveband of 625-774 nm, the spectral feature of Reflectance and the samples modeling proportion of 3:1. The number of principal components was 11, the training set accuracy rate was 96.97%, the testing set accuracy rate was 98.18%. The Second one was the model with the waveband of 775-924 nm, the spectral feature of the 1st derivative of Reflectance and the samples modeling proportion of 3:1. The number of principal components was 9, and the training set accuracy rate was 95.15%, and the testing set accuracy rate was 98.18%. |
PDF Full Text Request